Trends · 研究趋势
数据来自 Hugging Face 论文的 AI 提取关键词,按月统计研究方向的增长与热度。
Hongyuan Adam Lu, Z. L. Victor Wei, Qun Zhang +28 authors
Looped World Models introduce iterative latent state refinement through shared transformer blocks, achieving 100x parameter efficiency while adapting computational depth to prediction complexity.
DeepSeek-AI, Aixin Liu, Aoxue Mei +261 authors
DeepSeek-V3.2 introduces DeepSeek Sparse Attention and a scalable reinforcement learning framework, achieving superior reasoning and performance compared to GPT-5 and Gemini-3.0-Pro in complex reasoning tasks.
Zhiliang Peng, Jianwei Yu, Wenhui Wang +10 authors
VibeVoice synthesizes long-form multi-speaker speech using next-token diffusion and a highly efficient continuous speech tokenizer, achieving superior performance and fidelity.
Zhiyu Li, Shichao Song, Chenyang Xi +36 authors
MemOS, a memory operating system for Large Language Models, addresses memory management challenges by unifying plaintext, activation-based, and parameter-level memories, enabling efficient storage, retrieval, and continual learning.
Boqiang Zhang, Lei Ke, Ruihan Yang +5 authors
Penguin-VL demonstrates that text-only initialized vision encoders can achieve superior performance in multimodal understanding tasks compared to traditional contrastive pretraining methods, enabling efficient deployment on resource-constrained devices.
Yuling Shi, Chaoxiang Xie, Zhensu Sun +7 authors
Multimodal large language models can effectively understand source code when represented as compressed images, achieving significant token reduction while maintaining or improving performance on code comprehension tasks.
Shuming Ma, Hongyu Wang, Shaohan Huang +5 authors
BitNet b1.58 2B4T, a 1-bit Large Language Model with 2 billion parameters, matches the performance of full-precision models while improving computational efficiency.
Ling-Team, Ang Li, Ben Liu +139 authors
Ling 2.0, a reasoning-oriented language model series, achieves high efficiency and accuracy through a Mixture-of-Experts paradigm, sparse activation, and innovative training techniques.
Vishesh Tripathi, Abhay Kumar
Grouped Query Experts (GQE) improves Transformer efficiency by selectively activating query heads based on token content while maintaining key-value cache benefits of grouped-query attention.
Jingwei Zuo, Maksim Velikanov, Ilyas Chahed +24 authors
Falcon-H1, a new series of large language models with a hybrid architecture combining Transformer-based attention and State Space Models, achieves state-of-the-art performance and efficiency across various tasks and sizes.
Chin-Yang Lin, Cheng Sun, Fu-En Yang +3 authors
LongSplat improves novel view synthesis from long videos with irregular motion through joint optimization, robust pose estimation, and efficient anchor formation.
Mihir Prabhudesai, Russell Mendonca, Zheyang Qin +2 authors
Utilizing pre-trained reward models to adapt video diffusion models with gradient-based feedback enhances efficiency and performance compared to gradient-free methods.
Wei Chen, Zhiyuan Li, Zhen Guo +1 authors
An on-device Planner-Action framework, utilizing Phi-3 Mini for planning and Octopus for execution, employs model fine-tuning and multi-LoRA training to enhance efficiency and multi-domain handling on resource-constrained devices.
Chi Han, Qifan Wang, Wenhan Xiong +3 authors
LM-Infinite addresses length generalization in Transformer-based LLMs using a simple $\Lambda$-shaped attention mask, achieving fluency and quality up to 32k tokens without retraining.
Zhengqing Yuan, Zhaoxu Li, Lichao Sun
TinyGPT-V, built on Phi-2 with vision modules from BLIP-2 or CLIP, offers high performance with low computational requirements, enabling efficient multimodal large language modeling.
Gregor Bachmann, Sotiris Anagnostidis, Thomas Hofmann
MLPs achieve competitive performance on vision tasks with large-scale pre-training, challenging the narrative that inductive bias is necessary for high accuracy.
北京市昌平区探索星信息技术及软件开发工作室
京ICP备2026059466号